What causes the accuracy gap between spherical and perspective panorama SfM?

Hi, thank you for providing the panorama_sfm.py example.

The script notes that the virtual-rig approach is “generally more accurate but slower” than direct equirectangular reconstruction. Could you clarify the main reason for this accuracy gap?

I currently have two hypotheses:

1. Feature extraction and matching may be less robust on equirectangular images because of spatially varying distortion, especially near the poles.
2. During bundle adjustment, reprojection errors may be measured in equirectangular pixel space without accounting for latitude-dependent distortion, which could overweight errors near the poles.

I would appreciate any insight/numbers into which of these, or other factors, matter most in practice.

Thank you!